Simulation is pivotal in evaluating the performance of autonomous driving systems due to the advantages in efficiency and cost compared to on-road testing. Realistic multi-agent behavior~(e.g., interactive and long-term) is needed to narrow the gap between the simulation and the reality. The existing work has the following shortcomings in achieving this goal:~(1) log replay offers realistic scenarios but leads to unrealistic collisions due to lacking dynamic interactions, and~(2) model-based and learning-based solutions encourage interactions but often deviate from real-world data in long horizons. In this work, we propose LitSim, a long-term interactive simulation approach that maximizes realism while avoiding unrealistic collisions. Specifically, we replay the log for most scenarios and intervene only when LitSim predicts unrealistic conflicts. We then encourage interactions among the agents and resolve the conflicts, thereby reducing the likelihood of unrealistic collisions. We train and validate our model on the real-world dataset NGSIM, and the experimental results demonstrate that LitSim outperforms the current popular approaches in realism and reactivity.

翻译：仿真因其在效率和成本上相较于道路测试的优势，在评估自动驾驶系统性能中起着关键作用。为了缩小仿真与现实之间的差距，需要实现逼真的多智能体行为（例如交互性与长期性）。现有工作在达成此目标方面存在以下不足：（1）日志回放可提供真实场景，但因缺乏动态交互而导致不真实的碰撞；（2）基于模型和基于学习的方法虽鼓励交互，但在长时间跨度内常偏离真实数据。本文提出LitSim——一种长期交互式仿真方法，旨在最大化真实性的同时避免不真实的碰撞。具体而言，我们回放大多数场景的日志，仅当LitSim预测到不真实的冲突时进行干预。随后，我们鼓励智能体间的交互并解决冲突，从而降低不真实碰撞的可能性。我们在真实世界数据集NGSIM上训练并验证了模型，实验结果表明LitSim在真实性和反应性上均优于当前主流方法。